The next generation of LLM architecture; Project X is a neuro-symbolic cryptographic AI engine that mathematically matches fragmented solutions to complex problems—across classification boundaries, organisational silos, and entire industries. Grounded with Surprisal and Knowledge Graphs, secured at the hardware level with Trusted Execution Environments.
Organisations cannot see what they already have. Siloed data, incompatible taxonomies, and classification boundaries mean that critical solutions remain buried—while identical problems are re-commissioned at enormous cost.
A defence "retention" challenge and a corporate "attrition" problem are semantically identical—but current systems never link them. Procurement cycles restart from zero, duplicating spend on problems already solved elsewhere in the ecosystem.
Large organisations with siloed business units unknowingly commission parallel workstreams. Useful IP sits in dormant repositories while new contracts are issued for near-identical requirements.
Classification barriers prevent cross-domain visibility. Solutions on the high-side cannot be discovered from the low-side, and vice versa. Innovation stalls at the boundary.
A proprietary architecture that fuses the intuition of neural networks with the rigour of symbolic reasoning—grounded through information-theoretic Surprisal and dynamic Knowledge Graphs, all running inside Trusted Execution Environments for data confidentiality even during computation.
Vector embeddings capture fuzzy semantic similarity. A defence "personnel retention framework" and a corporate "employee churn model" are recognised as conceptually identical—no shared keywords required.
A dynamic knowledge graph constrains and structures neural output. Ontological reasoning ensures matches are logically valid and auditable—not probabilistic guesses.
Information-theoretic Surprisal grounds the LLM outputs, quantifying how unexpected a semantic connection is within the graph topology. This filters noise and ensures the engine surfaces genuinely novel, high-value matches rather than trivially obvious ones.
A combinatorial optimisation process that identifies subsets of partial solutions whose union satisfies a requirement. Solution A (60% fit) combined with Solution B (30% fit) covers a requirement—surfacing combinations that standard vector-search categorically misses. This is an NP-Hard problem space.
All matching computation executes inside hardware-level TEEs (e.g., Intel SGX, ARM TrustZone). Data remains encrypted in memory—not even the host infrastructure operator can access it during processing.
Classified and sensitive node data is abstracted before matching. High-side problems can be matched against low-side solutions without exposing the underlying data to either party.
Every match produces an auditable reasoning chain. Critical for government procurement audit trails, regulatory compliance, and building institutional trust in AI-assisted decisions.
Early benchmarks from our TRL 4 prototype tested against structured problem-solution datasets. These represent controlled-environment results; production performance at scale is the next phase of validation.
Correctly identified relevant solutions across synonym-heavy, jargon-different problem descriptions where zero keywords overlapped.
Successfully identified multi-solution compositions (2–4 partial solutions combining to cover a requirement) that single-vector search returned zero results for.
Of all matches surfaced, the percentage that produced a complete, human-readable reasoning chain traceable back through the knowledge graph.
Improvement in relevant matches surfaced compared to traditional keyword and TF-IDF based search across the same dataset, with a 3x improvement in recall for multi-vendor solution coverage.
Zero plaintext data exposure events during TEE-secured matching operations. All computation completed within encrypted memory enclaves.
Average time to return ranked composite matches against a 10,000-node graph, including TEE attestation overhead. Target for production: <200ms at 1M+ nodes.
All benchmarks derived from controlled prototype testing (TRL 4) against curated datasets. These figures are indicative of architectural capability, not production-scale guarantees. Independent validation is a stated objective of our next development phase.
Built from day one for deployment in the most sensitive environments. Our protocol-first architecture assumes hostile infrastructure and protects data at every layer—from storage through to computation. Meeting NCSC cloud security principles by design.
Matching computation runs inside hardware-isolated enclaves (Intel SGX / ARM TrustZone compatible). Data is decrypted only inside the enclave—the host OS, hypervisor, and even physical access to the server cannot compromise it. Remote attestation cryptographically proves code integrity before any data enters the enclave.
Before data enters the matching engine, a semantic abstraction layer strips identifying details while preserving the mathematical properties needed for matching. Classified problem statements can be compared against unclassified solution databases without exposing the underlying requirement.
Our decentralised protocol architecture supports air-gapped deployment. Secure instances can operate entirely self-hosted on classified networks, with controlled federation to lower-classification instances via the abstraction layer. No SaaS dependency. No data leaves the boundary.
Every match, every reasoning step, every data access event is logged in an immutable audit chain. Designed to meet government audit requirements including NCSC guidelines and MOD information assurance standards. Full XAI output means no "black box" decisions.
The most powerful emergent property: as the knowledge graph grows, connections between seemingly unrelated industries surface organically. Automotive sensor IP solves medical imaging problems. Aerospace materials address energy storage challenges. The graph itself becomes the moat.
Match capability gaps to existing solutions across classification boundaries. Reduce procurement duplication across MOD, DASA, and Five Eyes partner ecosystems.
Our initial route-to-market. Procurement and innovation units managing complex, high-security requirements can verify whether a capability already exists before commissioning new work. End users include requirements engineers, technical analysts, and procurement leads.
Connect fragmented clinical research, drug discovery pipelines, and MedTech solutions. Identify where existing approved compounds or devices address new therapeutic targets.
Eliminate duplication across siloed business units. Discover internal IP synergies before commissioning external work. Reduce R&D spend by identifying existing partial solutions. The core product is sector-agnostic and applies to any R&D-intensive enterprise with siloed solution repositories.
Connect space-qualified component databases to terrestrial applications. Match launch system requirements to advanced manufacturing capabilities across the UK supply chain.
Every problem and solution added to the graph increases the value for all participants. Cross-sector links compound exponentially. This unified data structure isn't just a feature—it's a proprietary, self-reinforcing data asset that grows more defensible with every node.
Problem statements and solution descriptions are ingested through the abstraction layer. Sensitive details are stripped; mathematical properties preserved. All processing inside TEE enclaves.
Neural embeddings capture semantic meaning. The symbolic layer maps them onto the knowledge graph with ontological constraints, creating structured, queryable nodes.
The engine runs a combinatorial optimisation process, identifying direct matches and composite combinations. Solution A + Solution B + Solution C may together cover a requirement that none addresses alone.
Every match returns a full reasoning chain: why these solutions, which properties aligned, what coverage gaps remain. Auditable. Verifiable. Human-readable.
We're building the intelligence layer that connects the world's fragmented innovation ecosystem. If you work in defence, government procurement, enterprise R&D, or deep tech—we want to hear from you.